DeepSeek founder says Nvidia’s CUDA moat is ‘rapidly disintegrating’

In a leaked transcript from a nearly four-hour investor meeting, DeepSeek founder Liang Wenfeng declared that Nvidia's ↘ 1.47% CUDA ecosystem monopoly is "rapidly disintegrating" and that China's AI chip industry faces a "historic opportunity" to replace foreign hardware, according to the record published by Tencent Tech in late July 2026.news.futunn+2

Compute as the Only Gap

Liang told investors that all differences between Chinese and American AI capabilities — including talent, model performance, and applications — trace back to a single variable: compute resources. He described computing capacity as DeepSeek's biggest constraint, noting the company currently operates roughly 20,000 Nvidia H-series equivalent chips but cannot procure enough hardware to train the largest frontier models.hellochinatech+2

"Spending 20 billion renminbi this year would mean our procurement team did an exceptional job. It is extremely difficult to spend that much. You can't buy that many chips, and the prices are high," Liang said, according to the transcript.hellochinatech

He predicted that within one year, real-world deployment would overturn market perceptions that China's domestic chip ecosystem is "unusable or difficult to use." Liang said DeepSeek's V3 model was trained on Nvidia chips but already bypassed Nvidia's software stack, using an internally built high-level compiler called TileLang instead of relying on CUDA.fredgao+2

Huawei Partnership and Domestic Chips

Liang was unusually direct about DeepSeek's collaboration with Huawei, stating that Huawei's 950 super-node "can fully replace Nvidia's GB200 and GB300 in performance and price," with four Huawei chips matching one Nvidia chip in capacity. He acknowledged that the remaining gap is "fourfold plus two years" on the chips themselves, with insufficient manufacturing capacity as the primary bottleneck.hellochinatech

Separately, Reuters reported in early July that DeepSeek is developing its own inference chip to further reduce reliance on both Nvidia and Huawei hardware.reuters

AGI Over Profit

Liang framed AGI as DeepSeek's singular goal, describing all commercial activity — consumer apps, enterprise APIs, and revenue — as byproducts of that pursuit. He outlined a roadmap progressing from language models through chain-of-thought reasoning and agents to continuous learning, which he called the next critical breakthrough before reaching artificial general intelligence.fredgao+1

The company plans to keep its strongest models open-source, with Liang dismissing closed-source strategies by pointing to ByteDance as an example with no visible benefit from proprietary models. On pricing, he said DeepSeek targets a ten-month payback period on hardware purchases — a margin he called "reasonable" but well below profit-maximizing levels.hellochinatech

"As long as I can keep the team stable, we will succeed at AGI. It is that simple," Liang said, identifying team retention — not capital or resources — as DeepSeek's only true internal risk.hellochinatech